电瓷避雷器Issue(3)：157-166,10.DOI:10.16188/j.isa.1003-8337.2024.03.020
改进Deeplab V3+在复合绝缘子红外热像分割中的应用
Application of Improved Deeplab V3+in Infrared Thermal Image Segmentation of Composite Insulators
摘要
Abstract
Aiming at the problems of low precision and large model of the existing segmentation algorithm used for uav infrared inspection of composite insulators,a real-time segmentation method based on im-proved Deeplab V3+for composite insulators infrared thermal image was proposed.Firstly,the revised MobileNet V3 is used as the Deeplab V3+backbone network to reduce the model complexity,and the spatially separable convolutional is adopted for the convolutional layer in the atrous spatial pyramid poo-ling to improve the inference speed of the network;secondly,in order to balance the missing feature in-formation and model accuracy degradation in the decoding network brought by the reduced amount of model parameters,Enhance insulator effective feature extraction by embedding the Enhanced Channel At-tention(ECA-Net)module in the MobileNet V3.Finally,PointRend is added to optimize the distribu-tion of sampling points and make the edge segmentation more accurate.By comparison with mainstream algorithms,the average cross-merge ratio of the optimized network is increased by 1.24％,and it can segment finer insulators from the IR thermal images,and the number of network parameters is only 9.2％of the original model,and the speed of segmentation resolution 512 × 512 IR thermal images reaches 56 fps,which provides an accurate and effective pre-processing means for the subsequent real-time defect detection of composite insulators.关键词
复合绝缘子/红外热像/实时语义分割/Deeplab V3+模型Key words
composite insulators/infrared images/real-time semantic segmentation/Deeplab V3+引用本文复制引用
黄靖,林杨,林朝晖,陈斌艺..改进Deeplab V3+在复合绝缘子红外热像分割中的应用[J].电瓷避雷器,2024,(3):157-166,10.基金项目
福建省科技厅工业引导性(重点)项目(编号:2021H0024) （重点）
国网福建省电力有限公司科技项目(编号:52130420001N).Project supported by Fujian Provincial Department of Science and Technology Industrial Guidance(Key)Project(No.2021H0024) （编号:52130420001N）
Science and Technology Program of State Grid Fujian Electric Power Co.,Ltd.,(No.52130420001N). （No.52130420001N）
